Can't use tf.timestamp() from within @tf.function with XLA / jit_compile=True

I would like to use tf.timestamp() when it is available (eager mode and graph mode without XLA), and use 0. (or a better fallback if there is one) when it is not available (with XLA; @tf.function(jit_compile=True)).

I tried this:

def tf_timestamp_or_zero():
    try:
        return tf.timestamp()
    except tf.python.framework.errors_impl.InvalidArgumentError:
        return 0.

#@tf.function
@tf.function(jit_compile=True)
def __call__(self):
    #...
    #t = tf.timestamp() # same error as below
    t = tf_timestamp_or_zero() # error, see below
    #...

Error:

tensorflow.python.framework.errors_impl.InvalidArgumentError:
Detected unsupported operations when trying to compile graph
_inference___call___5559[
    _XlaMustCompile=true,
    config_proto=3175580994766145631,
    executor_type=11160318154034397263
] on XLA_CPU_JIT:
Timestamp (No registered 'Timestamp' OpKernel for XLA_CPU_JIT devices
compatible with node {{node Timestamp}}){{node Timestamp}}

posted about 1 year ago

CC BY-SA 4.0

daniel_s‭

16 reputation 3 0 3 5